ebook img

the quantum adversary method and classical formula size lower bounds PDF

34 Pages·2006·0.33 MB·English
by  
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview the quantum adversary method and classical formula size lower bounds

comput. complex.15(2006),163–196 (cid:13)c Birkh¨auserVerlag,Basel2006 1016-3328/06/020163–34 computational complexity DOI10.1007/s00037-006-0212-7 THE QUANTUM ADVERSARY METHOD AND CLASSICAL FORMULA SIZE LOWER BOUNDS Sophie Laplante, Troy Lee, and Mario Szegedy Abstract. We introduce two new complexity measures for Boolean functions, which we name sumPI and maxPI. The quantity sumPI has been emerging through a line of research on quantum query complex- ity lower bounds via the so-called quantum adversary method (Ambai- nis 2002, 2003; Barnum et al. 2003; Laplante & Magniez 2004; Zhang 2005),culminatinginSˇpalek&Szegedy(2005)withtherealizationthat these many different formulations are in fact equivalent. Given that sumPI turns out to be such a robust invariant of a function, we begin to investigate this quantity in its own right and see that it also has ap- plications to classical complexity theory. As a surprising application we show that sumPI2(f) is a lower bound on the formula size, and even, up to a constant multiplicative factor, the probabilistic formula size of f. Weshowthatseveralformulasizelowerboundsintheliterature,specif- icallyKhrapchenkoanditsextensions(Khrapchenko1971;Koutsoupias 1993), including a key lemma of H˚astad (1998), are in fact special cases of our method. The second quantity we introduce, maxPI(f), is always at least as large as sumPI(f), and is derived from sumPI in such a way that maxPI2(f) remains a lower bound on formula size. Our main re- sult is proven via a combinatorial lemma which relates the square of thespectralnormofamatrixtothesquaresofthespectralnormsofits submatrices. Thegeneralityofthislemmaimpliesthatourmethodscan alsobeusedtolower-boundthecommunicationcomplexityofrelations, and a related combinatorial quantity, the rectangle partition number. To exhibit the strengths and weaknesses of our methods, we look at the sumPI and maxPI complexity of a few examples, including the recursive majority of three function, a function defined by Ambainis (2003), and the collision problem. Keywords. Lower bounds, quantum computing, adversary method, formula size, communication complexity. Subject classification. 68Q17, 68Q30. 164 Laplante, Lee & Szegedy cc 15 (2006) 1. Introduction A central and longstanding open problem in complexity theory is to prove superlinear lower bounds for the circuit size of an explicit Boolean function. While this seems quite difficult, a modest amount of success has been achieved in the weaker model of formula size, a formula being simply a circuit where every gate has fan-out at most one. The current best formula size lower bound for an explicit function is n3−o(1) by H˚astad (1998). Inthispaperweshowthatpartoftherichtheorydevelopedaroundproving lower bounds on quantum query complexity, namely the so-called quantum adversaryargument,canbebroughttobearonformulasizelowerbounds. This adds to the growing list of examples of how studying quantum computing has ledtonewresultsinclassicalcomplexity, includingAaronson(2004);Kerenidis & Wolf (2004); Laplante & Magniez (2004); Sen & Venkatesh (2001), to cite a few. The roots of the quantum adversary argument can be traced to the hybrid argumentofBennettetal.(1997),whouseittoshowanΩ(√n)lowerboundon quantumsearch. Ambainisdevelopedamoresophisticatedadversaryargument (Ambainis 2002) and later improved this method to the full-strength quantum adversaryargument(Ambainis2003). FurthergeneralizationsincludeBarnum, Saks & Szegedy (2003) with their spectral method and Zhang (2005) with his strong adversary method. Laplante & Magniez (2004) use Kolmogorov com- plexity to capture the adversary argumentin terms ofa minimization problem. ThislineofresearchculminatesinrecentworkofSˇpalek&Szegedy(2005)who show that in fact all the methods of Ambainis (2003); Barnum et al. (2003); Laplante & Magniez (2004); Zhang (2005) are equivalent. The fact that the quantum adversary argument has so many equivalent definitions indicates that it is a natural combinatorial property of Boolean functions which is worth investigating on its own. We give this quantity its own name, sumPI, and adopt the following primal formulation of the method, fromLaplante&Magniez(2004);Sˇpalek&Szegedy(2005). LettingS 0,1 n ⊆{ } and f :S 0,1 be a Boolean function we set →{ } 1 (1.1) sumPI(f)=min max , p x,y p (i)p (i) f(x)6=f(y) i:xi6=yi x y P p wherep= p :x S isafamilyofprobabilitydistributionsontheindices[n]. x { ∈ } If Q (f) is the two-sided error quantum query complexity of f then Q (f) = ǫ ǫ Ω(sumPI(f)). We show further that sumPI2(f) is a lower bound on the formula size of f. Moreover, sumPI2(f) generalizes several formula size lower bounds in cc 15 (2006) The quantum adversary method and formula size 165 the literature, specifically Khrapchenko and its extensions (Khrapchenko 1971; Koutsoupias 1993), and a key lemma of H˚astad (1998) used on the way to proving the current best formula size lower bounds for an explicit function. We also introduce KI(f)= min max min (K(ix,α)+K(iy,α)), α∈Σ∗ x,y i:xi6=yi | | f(x)6=f(y) whereK istheprefix-freeKolmogorovcomplexity. Thisformulationarisesfrom thequantumandrandomizedlowerboundsofLaplante&Magniez(2004). This formulation is especially interesting because of the intuition that it provides. For example, it allows for a very simple proof that circuit depth d(f) KI(f), ≥ using the Karchmer–Wigderson characterization of circuit depth (Karchmer & Wigderson 1988). We define a quantity closely related to 2KI, which we call maxPI, by 1 (1.2) maxPI(f) =min max . p x,y max p (i)p (i) f(x)6=f(y) i:xi6=yi x y Notice that this is like sumPI but with the sum repplaced by a maximum. By definition, maxPI is larger than sumPI, but its square is still a lower bound on formula size. We prove our main results by transforming in two steps the problem of proving formula size lower bounds into a problem with a more combinatorial flavor which is easier to work with. First, we use the elegant characterization given by Karchmer & Wigderson (1988) of formula size in terms of the com- munication complexity of a relation. We then use the well-known property that a successful communication protocol partitions a relation into rectangles of a certain form. We then lower-bound the size of the smallet such rectangle partition. A sufficient condition for a measure to lower-bound the size of such a partition is that it is subadditive on disjoint rectangles. Our main lemma shows that the spectral norm squared of a matrix A is such a measure. We look at several concrete problems to illustrate the strengths and weak- nesses of our methods. We study the heighth recursive majority of three prob- lem, R-MAJh, and show that Q (R-MAJh)=Ω(2h) and a lower bound of 4h for 3 ǫ 3 theformulasize. WealsolookatafunctiondefinedbyAmbainis(2003)tosep- aratethequantumquerycomplexityofafunctionfromtheboundgivenbythe polynomial method (Beals et al. 2001). This function gives an example where sumPI2 cangive somethingmuchbetterthanKhraphchenko’sbound. Fortotal functions, maxPI and sumPI are polynomially related; however, we give an ex- ampleofapartialfunctionf,namelythecollisionproblem,wheresumPI(f)=2 166 Laplante, Lee & Szegedy cc 15 (2006) and maxPI(f) = Θ(√n). This example shows that in general maxPI is not a lower bound on quantum query complexity, as for the collision problem maxPI(f) Q (f)=Θ(n1/3) (Aaronson & Shi 2004; Brassard et al. 1997). ǫ ≫ 1.1. Organization. In Section 2, we give the definitions, results, and nota- tion that we use throughout the paper, and introduce the quantities sumPI, maxPI, and KI. In Section 3 we prove some properties of sumPI and maxPI. In Section4,weshowhowsumPIandmaxPIgiverisetoformulasizelowerbounds, for deterministic and probabilistic formula size. In Section 5, we compare our new methods with previous methods in formula size complexity. In Section 6, we investigate the limits of our and other formula lower bound methods. Fi- nally, in Section 7 we apply our techniques to some concrete problems. 2. Preliminaries We use standard notation such as [n] = 1,...,n , S for the cardinality of a { } | | set S, and all logarithms are base 2. Hamming distance is written d . H 2.1. ComplexitymeasuresofBooleanfunctions. Weusestandardmea- sures of Boolean functions, such as sensitivity and certificate complexity. We briefly recall these here; see Buhrman & Wolf (2002) for more details. For a set S 0,1 n and Boolean function f : S 0,1 , the sensitivity of f on ⊆ { } → { } input x is the number of positions i [n] such that changing the value of x ∈ in position i changes the function value. The zero-sensitivity, written s (f), is 0 the maximum over x f−1(0) of the sensitivity of f on x. The one-sensitivity, ∈ s (f), is defined analogously. The maximum of s (f),s (f) is the sensitiv- 1 0 1 ity of f, written s(f). For block sensitivity, one considers when the function changes not just by flipping one bit but by flipping a set (or block) of bits. A block is sensitive on x if flipping all the bits in the block changes the value of the function. The block sensitivity of f on input x is the maximum number of disjoint sensitive blocks for x. The block sensitivity of f, written bs(f), is the maximum over all inputs x of the block sensitivity of f on x. A certificate for f on input x S is a subset I [n] such that for any y ∈ ⊆ satisfying y = x for all i I it must be the case that f(y) = f(x). The zero- i i ∈ certificate complexity of f, written C (f), is the maximum over all x f−1(0) 0 ∈ of the minimum size certificate of x. Similarly, the one-certificate complexity of f, written C (f), is the maximum over all x f−1(1) of the minimum size 1 ∈ certificateofx. ThemaximumofC (f),C (f)isthecertificate complexity off, 1 0 written C(f). cc 15 (2006) The quantum adversary method and formula size 167 2.2. Linear algebra. For a matrix A (respectively, vector v) we write AT (resp. vT)forthetransposeofA, andA∗ (resp. v∗)fortheconjugatetranspose ofA. FortwomatricesA,B weletA B betheHadamardproductofAandB, ◦ that is, (A B)[x,y] =A[x,y]B[x,y]. We write A B if A is entrywise greater ◦ ≥ than B, and A B when A B is positive semidefinite, that is, if A B is (cid:23) − − Hermitian and vT(A B)v 0 for all vectors v. We let rk(A) denote the rank − ≥ of the matrix A. We will use the notation Entrysum(A) for A[i,j]. i,j We will make extensive use of the spectral norm, denoted A . For a 2 P k k matrix A, A = √λ:λ is the largest eigenvalue of A∗A . 2 k k { } For a vector v, we let v be the ℓ norm of v. 2 | | We will also make use of some other matrix norms. The maximum absolute columnsumnorm,written A ,isdefinedas A =max A[i,j],andthe k k1 k k1 j i| | maximum absolute row sum norm, written A , is A =max A[i,j]. k k∞ k k∞ P i j| | The Frobenius norm A = A[i,j]2 is the ℓ norm of A thought of as k kF i,j 2 P a long vector. q P We collect a few facts about the spectral norm. These can be found in, for example, Horn & Johnson (1999). Proposition 2.1. Let A be an arbitrary m by n matrix. Then u∗Av (i) A =max| |. 2 k k u,v u v | || | (ii) A 2 A A . k k2 ≤k k1k k∞ (iii) For nonnegative matrices A,B, if A B then A B . 2 2 ≤ k k ≤k k 2.3. Deterministic and probabilistic formulae. A Boolean formula over thestandardbasis , , isabinarytreewhereeachinternalnodeislabeled {∨ ∧ ¬} with or , and each leaf is labeled with a literal, that is, a Boolean variable ∨ ∧ or its negation. The size of a formula is its number of leaves. We naturally identify a formula with the function it computes. Definition 2.2. Let f : 0,1 n 0,1 be a Boolean function. The formula size of f, denoted L(f), is{the}siz→e o{f the}smallest formula which computes f. The formula depth of f, denoted d(f), is the minimum depth of a formula computing f. 168 Laplante, Lee & Szegedy cc 15 (2006) It is clear that L(f) 2d(f); that in fact the opposite inequality d(f) O(logL(f)) also holds is ≤a nontrivial result due to Spira (1971). ≤ We will also consider probabilistic formulae, that is, a probability distribu- tion over deterministic formulae. We take a worst-case notion of the size of a probabilistic formula. This model of formula size has been studied in the series ofworksBoppana(1989);Dubiner&Zwick(1997);Valiant(1984)whichinves- tigate constructing efficient deterministic monotone formulae for the majority function by amplifying the success probability of probabilistic formulae. The interested reader can also compare our definition with two different models of probabilistic formula size considered in Klauck (2004). Definition 2.3. Let f be a set of functions with f : S 0,1 for j j∈J j { } → { } each j J. For a function f : S 0,1 , we say that f is ǫ-approximated by ∈ → { } f if there is a probability distribution α = α over J such that for j j∈J j j∈J { } { } every x S, ∈ Pr [f(x)=f (x)] 1 ǫ. α j ≥ − In particular, if max L(f ) s, then we say that f is ǫ-approximated by j j formulas of size s, denoted Lǫ≤(f) s. ≤ Note that even if a function depends on all its variables, it is possible that the probabilistic formula size is less than the number of variables. 2.4. Communication complexity of relations. Karchmer & Wigderson (1988)giveanelegantcharacterizationofformulasizeintermsofacommunica- tion game. We will use this formulation in our proofs. This has the advantage of letting us work in the more general setting of communication complexity of relations and enabling us to use the combinatorial tools of communication complexity. We now describe the setting. Let X,Y,Z be finite sets, and R X Y Z. In the communication game ⊆ × × for R, Alice is given some x X, Bob is given some y Y and their goal is to ∈ ∈ find some z Z such that (x,y,z) R, if such a z exists. A communication ∈ ∈ protocol is a binary tree where each internal node v is labeled by either a function a : X 0,1 or b : Y 0,1 describing either Alice’s or Bob’s v v → { } → { } message at that node, and where each leaf is labeled with an element z Z. A ∈ communication protocol computes R if for all (x,y) X Y walking down the ∈ × tree according to a ,b leads to a leaf labeled with z such that (x,y,z) R, v v provided such a z exists. The communication cost D(R) of R is the he∈ight of the smallest communication protocol computing R. The protocol partition number CP(R) is the number of leaves in the smallest communication protocol computing R. cc 15 (2006) The quantum adversary method and formula size 169 Definition 2.4. With any Boolean function f we associate the relation R = (x,y,i) :f(x)=0, f(y)=1, x =y . f i i { 6 } Theorem 2.5 (Karchmer–Wigderson). For any Boolean function f, L(f) = CP(R ) and d(f)=D(R ). f f An advantage of the communication complexity approach to formula size is that we can use the powerful combinatorial tools available for communica- tion complexity lower bounds. At the heart of this approach lies the idea of combinatorial rectangles. A combinatorial rectangle is simply a set S X Y ⊆ × which can be expressed as S = X′ Y′ for some X′ X and Y′ Y. We × ⊆ ⊆ say that a set S X Y is monochromatic with respect to the relation R if ⊆ × there is a z Z such that (x,y,z) R for all (x,y) S. It can be shown ∈ ∈ ∈ that the leaves of a successful communication protocol for R form a disjoint covering of X Y by rectangles monochromatic with respect to R. We let × CD(R)bethesizeofthesmallestdisjointcoveringofX Y bymonochromatic × rectangles. It follows that CD(R) CP(R). For more information on commu- ≤ nication complexity and proofs of the above results, we suggest Kushilevitz & Nisan (1997). 2.5. sumPIandthequantumadversarymethod. Knowledgeofquantum computing is not needed for reading this paper; for completeness, however, we briefly sketch the quantum query model. More background on quantum query complexity and quantum computing in general can be found in Buhrman & Wolf (2002); Nielsen & Chuang (2000). As with the classical counterpart, in the quantum query model we wish to compute some function f : S 0,1 , where S Σn, and we access the → { } ⊆ input through queries. The complexity of f is the number of queries needed to compute f. Unlike the classical case, however, we can now make queries in superposition. Formally, a query O corresponds to the unitary transformation O : i,b,z i,b x ,z , i | i 7→| ⊕ i where i [n], b 0,1 , and z represents the workspace. A t-query quantum ∈ ∈ { } algorithm A has the form A = U OU O OU OU , where the U are fixed t t−1 1 0 k ··· unitary transformations independent of the input x. The computation begins in the state 0 , and the result of the computation A is the observation of the | i rightmost bit of A0 . We say that A ǫ-approximates f if the observation of | i the rightmost bit of A0 is equal to f(x) with probability at least 1 ǫ, for | i − 170 Laplante, Lee & Szegedy cc 15 (2006) every x. We denote by Q (f) the minimum query complexity of a quantum ǫ query algorithm which ǫ-approximates f. Alongwiththepolynomialmethod(Bealsetal.2001),oneofthemaintech- niques for showing lower bounds in quantum query complexity is the quantum adversarymethod(Ambainis2002,2003;Barnumet al.2003;Laplante&Mag- niez 2004; Zhang 2005). Recently, Sˇpalek & Szegedy (2005) have shown that all the strong versions of the quantum adversary method are equivalent, and further that these methods can be nicely characterized as primal and dual. We give the primal characterization as our principal definition of sumPI. Definition 2.6. Let S 0,1 n and f : S 0,1 be a Boolean function. Foreveryx S letp :[n⊆] { R}beaprobabili→tyd{istri}bution, thatis, p (i) 0 x x ∈ → ≥ and p (i) = 1. Let p = p : x S . We define the sum probability of i x { x ∈ } indices to be P 1 sumPI(f)=min max . p x,y p (i)p (i) f(x)6=f(y) i:xi6=yi x y P p We will also use two versions of the dual method, both a weight scheme and the spectral formulation. The most convenient weight scheme for us is the “probability scheme”, given in Lemma 4 of Laplante & Magniez (2004). Definition 2.7 (Probability scheme). Let S 0,1 n and f : S 0,1 ⊆ { } → { } be a Boolean function, and X = f−1(0), Y = f−1(1). Let q be a probability distribution on X Y, and p ,p be probability distributions on X,Y respec- A B × tively. Finally, let p′ : x X, i [n] and p′ : y Y, i [n] be families { x,i ∈ ∈ } { y,i ∈ ∈ } of probability distributions on X, Y respectively. Assume that q(x,y) = 0 when f(x) = f(y). Let P range over all possible tuples (q,p ,p , p′ ) of A B { x,i}x,i distributions as above. Then p (x)p (y)p′ (y)p′ (x) A B x,i y,i PA(f)=max min . P x,y,i q q(x,y) q(x,y)6=0,xi6=yi We will also use the spectral adversary method. Definition 2.8 (Spectral adversary). Let S 0,1 n and f : S 0,1 be ⊆ { } → { } a Boolean function. Let X = f−1(0), Y = f−1(1). Let A = 0 be an arbitrary 6 X Y nonnegative matrix. For i [n], let A be the matrix i | |×| | ∈ 0 if x =y , A [x,y]= i i i A[x,y] if x =y . i i (cid:26) 6 cc 15 (2006) The quantum adversary method and formula size 171 Then A SA(f)=max k k2 . A maxi Ai 2 k k Notethatthespectraladversarymethodwasinitiallydefined(Barnumetal. 2003) for symmetric matrices over X Y. The above definition is equivalent: ∪ if A is a symmetric matrix over X Y satisfying the constraint A[x,y] = 0 ∪ when f(x) = f(y), then A is of the form A = 0 B for some matrix B over BT 0 X Y. Then the spectral norm of A is equal to that of B. Similarly, for any × (cid:2) (cid:3) X Y matrix A we can form a symmetrized version of A as above preserving × the spectral norm. We will often use the following theorem implicitly in taking the method most convenient for the particular bound we wish to demonstrate. Theorem 2.9 (Sˇpalek–Szegedy). Let n 1 be an integer, S 0,1 n and ≥ ⊆ { } f :S 0,1 . Then →{ } sumPI(f)=SA(f)=PA(f). 2.6. The KI and maxPI complexity measures. The definition of KI arises fromtheKolmogorovcomplexityadversarymethod(Laplante&Magniez2004). The Kolmogorov complexity C (x) of a string x, with respect to a universal U Turing machine U, is the length of the shortest program p such that U(p) =x. The complexity of x given y, denoted C(xy), is the length of the shortest | program p such that U( p,y ) = x. When U is such that the set of halting h i programs is a prefix-free (no string in the set is a prefix of another in the set), wewriteK (xy). Fromthispointonwards, wefixU andsimplywriteK(xy). U | | For more background on Kolmogorov complexity consult Li & Vit´anyi (1997). Definition 2.10. For S 0,1 n and f :S 0,1 , let ⊆{ } →{ } KI(f)= min max min (K(ix,α)+K(iy,α)). α∈{0,1}∗ x,y i:xi6=yi | | f(x)6=f(y) The advantage of using concepts based on Kolmogorov complexity is that theyoftennaturallycapturetheinformation-theoreticcontentoflowerbounds. Asanexampleofthis,wegiveasimpleproofthatKIisalowerboundoncircuit depth. Theorem 2.11. For any Boolean function f, KI(f) d(f). ≤ Proof. Let P be a protocol for R . Fix x,y with different values under f, f and let T be a transcript of the messages sent from A to B, on input x,y. A 172 Laplante, Lee & Szegedy cc 15 (2006) Similarly, let T be a transcript of the messages sent from B to A. Let i be B the output of the protocol, with x = y . To print i given x, simulate P using i i 6 x and T . To print i given y, simulate P using y and T . This shows that B A x,y : f(x) = f(y), i : x = y ,K(ix,α)+K(iy,α) T + T D(R ), i i A B f w∀here α is a6 descript∃ion of6A’s and B| ’s algorithm| s. ≤ | | | | ≤ (cid:3) Remark. A similar proof in fact shows that KI(f) 2N(R ), where N is the f ≤ nondeterministic communication complexity. Since the bound does not take advantage of interaction between the two players, in many cases we cannot hope to get optimal lower bounds using these techniques. An argument similar to that in Sˇpalek & Szegedy (2005) shows that 1 2KI(f) =Θ min max . p x,y max p (i)p (i) (cid:18) f(x)6=f(y) i:xi6=yi x y (cid:19) p Notice that the right hand side of the equation is identical to the definition of sumPI, except that the sum in the denominator is replaced by a maximum. This led us to define the complexity measure maxPI, in order to get stronger formula size lower bounds. Definition 2.12. Let S 0,1 n and f : S 0,1 . For every x S let p :[n] R be a probabili⊆ty{distr}ibution. Let p→={p :}x S . We defi∈ne the x x → { ∈ } maximum probability of indices to be 1 maxPI(f) =min max . p x,y max p (i)p (i) f(x)6=f(y) i:xi6=yi x y p It can be easily seen from the definitions that sumPI(f) maxPI(f) for ≤ any f. The following lemma is also straightforward from the definitions: Lemma 2.13. IfS′ S andf′ :S′ 0,1 isadomainrestrictionoff :S 0,1 to S′, then sum⊆PI(f′) sumP→I(f{) and}maxPI(f′) maxPI(f). → { } ≤ ≤ 3. Properties of sumPI and maxPI 3.1. Properties of sumPI. Although in general, as we shall see, sumPI gives weaker formula size lower bounds than maxPI, the measure sumPI has several nice properties which make it more convenient to use in practice. ThenextlemmashowsthatsumPIbehaveslikemostothercomplexitymea- sures with respect to composition of functions:

Description:
adversary argument, can be brought to bear on formula size lower bounds. This adds to the growing list of A. E. Andreev (1987). On a method for
See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.